Zhongji Innolight

Phase A — Understand the business

Lens 1 · Company Overview

Zhongji Innolight (中际旭创) is the world's largest maker of high-speed optical transceiver modules — the small pluggable devices that convert electrical signals to light and back, plugged into switches and AI servers to move data between thousands of GPUs at 800 Gbps and now 1.6 Tbps ``. In the AI-infrastructure stack it is a pure picks-and-shovels supplier: every large GPU cluster needs hundreds of thousands of these modules to lash accelerators into a fabric, and Innolight is the volume leader.

The corporate structure is unusual and important. The listed shell was Shandong Zhongji Electrical Equipment — a maker of electrical switchgear founded by Wang Weixiu in Longkou, Shandong (origins 1987) . In **2016 Wang paid ¥2.8bn (~$412m) to acquire Suzhou InnoLight**, the optical-module company founded by Tsinghua-educated returnee **Dr. Liu Sheng** — a deal worth nearly 5× the acquirer's total assets . The merged entity was renamed Zhongji Innolight; the legacy switchgear business is now an immaterial rump. Optical-communication transceiver modules were 97.95% of FY2025 revenue `` — this is effectively a single-product company.

Business model: design + high-volume manufacturing/assembly of optical modules. Innolight buys the two expensive active components — the laser/optical chips (EML, externally-modulated lasers) and the DSP (digital signal processor) — packages them with passives into a qualified module, and sells to switch OEMs and directly to AI-system buyers. Contracts are purchase-order driven, not take-or-pay; pricing resets each generation. The tell on demand: on the April 2026 institutional call, management said key customers have already placed orders for the entirety of 2026 and the company is preparing 2027 orders , and framed 2026 as the industry's shift "from an order-competition phase to a delivery-assurance era" .

Customers: heavily concentrated in US hyperscalers and Nvidia. Innolight holds >50% of Nvidia's 800G optical-module procurement / wallet share , and **overseas revenue is ~91–92% of total** — i.e. this is a Chinese national champion whose end-demand is almost entirely American AI capex. End buyers include Nvidia (as a system integrator buying modules for its reference designs), plus Google/Meta/Amazon/Microsoft directly or via their switch ODMs.

Suppliers: the structural vulnerability (Lens 2/3) — DSP from Broadcom & Marvell (US duopoly), EML lasers from Lumentum, Coherent and others (also largely US/Western) ``.

Competitors: fellow Chinese champion Eoptolink (300502.SZ), plus Western vertically-integrated players Coherent (COHR) and Lumentum (LITE), China's Accelink (002281.SZ), and Broadcom (which sells both components and increasingly CPO switches).

Lens 2 · Supply Chain — named stakeholders

The optical-module chain, end to end, with the actual names:

Upstream materials/wafers → component makers (the value & the chokepoints):

• DSP / retimer silicon: Broadcom and Marvell — a near-duopoly ``. This is the single most expensive die in an 800G/1.6T DSP-based module and Innolight does not make it. Chokepoint #1 — and it is a US/Taiwan-anchored one.
• EML lasers / optical chips: Lumentum, Coherent, plus II-VI-lineage and Japanese suppliers; some Chinese supply (e.g. Source Photonics, HG Genuine/HGTECH) is climbing but not yet at leading-edge volume . EML is described as "a physical supply shortage," not just a duopoly . Chokepoint #2 — a hard capacity constraint that gates the whole industry's unit growth.
• Foundry/wafer for silicon photonics: TSMC / GlobalFoundries for SiPh; relevant as the industry shifts toward silicon-photonics-based 1.6T/CPO.
• Passives, PCBs, connectors, gold-box packaging: broad Chinese supplier base — this is where Innolight's domestic cost advantage lives.

The company: Innolight designs and assembles the module, qualifies it with each customer, and runs high-volume manufacturing. Capacity expanded ~34% YoY in FY25, shipments +45%, utilisation rose from 74% (FY24) to 85% (FY25) ``.

Manufacturing geography (a tariff-hedge chokepoint of its own): Innolight runs Thailand as its primary base for North American shipments and has added a Mexico facility; engineering and the deep supply chain stay anchored in China ``. This China-designs/Thailand-ships structure is the workaround for US tariffs and any Entity-List risk — and is itself a dependency (a single overseas ramp executing flawlessly).

Downstream → end customer: Innolight → switch OEM/ODM (or directly) → Nvidia / Google / Meta / Amazon / Microsoft AI data centres. Single-source-style dependency runs both ways: Innolight needs Nvidia's volume; Nvidia needs Innolight's qualified capacity to ship systems on time.

Verdict on the chain: Innolight sits at the fattest-volume but thinnest-IP node. The two pieces of silicon that actually gate performance and capture the economics (DSP, EML) are upstream and mostly American. That is the whole bear case in one sentence (see Lens 13) — and the answer to "Chinese modules own 7 of the top 10 seats, so why do the chip companies still make the money?" ``.

Lens 3 · Competitive Advantages (moats)

What's genuinely durable:

1. Scale + yield + qualification lead. Innolight is #1 by volume and was first to complete Nvidia's 1.6T qualification ``. Customer qualification for a new module generation is a 6–12 month, design-locked process; being first in at volume with good yields is a real, if generation-limited, switching cost.
2. Cost position. Innolight + Eoptolink manufacture 60%+ of 800G modules at 20–25% lower prices than Western makers, and run ~20–22% net margins `` — Chinese labour/supply-chain cost plus genuine manufacturing competence. As they climb the speed ladder they compress Western competitors' margins.
3. Trusted-supplier status with Nvidia — ">50% wallet share" is partly relationship/reliability capital that a new entrant cannot instantly replicate ``.

Why the moat is shallow (this matters as much as the moat itself):

• No control of the two key components. Bargaining power over Broadcom/Marvell (DSP) and Lumentum/Coherent (EML) is weak — these are concentrated, supply-constrained suppliers selling into a hot market. Innolight is a price-taker on its most expensive inputs.
• Customer power is enormous. Nvidia + a handful of hyperscalers are the demand; they can (and do) qualify second sources (Eoptolink), push price, and — most dangerously — move the architecture (CPO/LPO/in-house) to disintermediate module assembly entirely.
• Switching costs reset every generation. The qual lead at 800G doesn't guarantee 1.6T or 3.2T; it must be re-won each cycle.

Net: a process/scale/cost moat that is real but generation-scoped and sandwiched between powerful suppliers and powerful customers. Not a toll bridge — more a very good factory in a very good location, which the landlord (Nvidia) is studying how to bypass.

Lens 4 · Segments (product & geography)

No segments.csv on the shelf (unavailable), so all figures are:

• By product: optical-communication transceiver modules = 97.95% of FY25 revenue (¥37.46bn); legacy electrical equipment is the immaterial remainder . Within modules, the mix is shifting hard toward high-speed: **800G is the FY25 volume engine; 1.6T began ramping with key customers in H2 2025 and is expected to grow over the next several quarters** . The product-mix shift up the speed curve is the margin story — gross margin expanded ~8pts to 42.61% in FY25 as 800G/1.6T share rose ``.
• By geography: overseas ~91–92% of revenue (~86% of shipments) `` — overwhelmingly North America. This is the defining feature: a Shenzhen-listed company whose P&L is an almost-pure bet on US hyperscaler capex, run through a Thailand/Mexico shipping structure to manage tariff/geopolitical risk.
• Trend & cause: accelerating, and the cause is singular — the AI training/inference buildout driving GPU-cluster interconnect demand. The risk embedded in the trend is its own concentration (one end-demand, one architecture, a few customers).

---

Phase B — Measure performance

Lens 5 · Earnings Result (latest prints)

FY2025 (reported ~March 2026):

• Revenue ¥38.24bn, +60.25% YoY (FY24 ¥23.86bn) ``.
• Net profit ¥10.8bn, +108.78% YoY — profit more than doubled ``.
• Gross margin 42.61%, +~8pts YoY; module GM expansion driven by 800G/1.6T mix ``.
• Q4 FY25 alone: revenue ¥13.235bn, net profit ¥3.665bn ``.
• USD frame: ~$5.6bn revenue / ~$1.6bn net profit for FY25 (≈ at 7.15 RMB/USD: $5.35bn / $1.51bn — note the modest gap vs. the cited USD figures; the conflict is flagged, not silently resolved).
• Dividend: ¥10 per 10 shares, ¥1.111bn total — a token payout (~0.23% yield), correctly subordinated to reinvestment in a hypergrowth business ``.

Q1 2026 (reported ~17 April 2026) — the blowout:

• Revenue ¥19.50bn, +192.12% YoY ``.
• Net profit ¥5.73bn, +262.28% YoY — a single quarter already exceeded full-year 2024 net profit ``.
• Driver: 800G still growing + 1.6T ramping; mix continued to lift gross and net margins ``.
• Demand signal: customers booked all of 2026, 2027 booking underway ``.

Balance-sheet / quality flags (web-only, low granularity): capacity +34% and utilisation 74%→85% imply the growth is real volume, not just price ``. The two items to watch that the web does not cleanly disclose: inventory and receivables (a module maker pre-buying scarce EML/DSP can balloon inventory; hyperscaler terms can stretch receivables). These cannot be verified web-only — explicit open item for a filings-grounded refresh.

Market reaction: the stock is a top performer on China's onshore benchmark, all-time high ¥1,368.50 on 18 June 2026 `` — the tape has rewarded every print. The reaction to the prints has compressed forward P/E even as the price rose, because earnings outran the multiple.

Lens 6 · Earnings Calls / management focus (sentiment trend)

No transcripts/ on the shelf; this is assembled from call read-outs ``. Management's narrative has escalated in confidence across the last several quarters:

• Tone: increasingly assertive. FY25 → Q1'26 commentary moved from "strong demand" to "customers have booked all of 2026, we're taking 2027 orders" and the framing that the bottleneck is now delivery, not order capture ("delivery-assurance era") ``.
• Recurring themes: high-speed mix (800G→1.6T) lifting margins; capacity/utilisation ramp; overseas (US) demand; LPO/CPO as something they're "actively deploying" while "balancing investment with market demand" ``.
• On the CPO threat (the key tell): management told analysts no cloud-service-provider customer plans large-scale CPO deployment in 2026–2027 `` — i.e. they are publicly dismissing the near-term disintermediation risk. Worth weighing as interested testimony.
• What they've stopped emphasising: generic "400G" / legacy-speed commentary — the story is now entirely leading-edge.

Sentiment trend = rising and almost promotional, consistent with a company at the peak of its cycle. The discipline to watch for is whether confidence tips into over-ordering of components and capacity into an eventual digestion.

Lens 7 · Comps

Peer table — Innolight vs. global optical-transceiver peers. Multiples are `` with date, or n/a. None are fabricated. A-share names trade at structurally different multiples than US names; cross-border P/E comparisons are directional, not precise.

Read: on forward P/E (~29×), Innolight is the cheapest name in its own peer group — below Coherent (~49×), Lumentum (~59×) and far below Eoptolink (~198× TTM) . That is the single most striking valuation fact: the volume leader with the best margins and the Nvidia wallet trades at the lowest forward multiple. Either the market is pricing in the 1.6T margin/share risk (CPO, in-house, EML constraints) — or there's a genuine mispricing. The TTM-vs-forward gap (66×→29×) encodes the market's own expectation of a **massive FY26 earnings step-up** that must actually land. *(Market-cap range is unusually wide across sources/dates — treat the USD figure as approximate; one normalized "fair-value premium" screen flagged the stock ~473% above its model fair value , which is a momentum-name artefact more than a verdict.)*

Lens 8 · Stock-Price Catalysts (what actually moves it)

Pattern over the AI cycle ``:

• Up: every AI-capex/GPU-demand upgrade and every earnings beat. FY25 and Q1'26 prints (triple-digit growth) drove the stock to triple-digit annual gains and a top spot on the onshore benchmark ; ATH ¥1,368.50 on 18 Jun 2026. Nvidia order/qual news (first to pass 1.6T) is a direct catalyst .
• Down: architecture-threat headlines are the kryptonite. The Jan 2026 DeepSeek "efficient-AI / less-capex" scare and recurring CPO/in-house-optics narratives are the sell-off triggers (the web record for this name is thin on the specific intraday DeepSeek drawdown, but the AI-capex-doubt complex is clearly the down-catalyst category). Nvidia's 2 March 2026 $4bn CPO investment into Coherent + Lumentum is exactly the kind of event that pressures the module-assembler thesis ``.
• What the market reacts to: (1) the level of hyperscaler/Nvidia capex, (2) Innolight's share of that capex (wallet-share-shift fears), (3) the architecture (does the module survive?). Earnings beats are necessary but the re-rating swings on items 2 and 3. This is a name that trades on "is the pluggable optical module still the answer in 2 years?" as much as on its own numbers.

---

Phase C — Judge people & books

Lens 9 · Management

• Dr. Liu Sheng — Chairman (since Aug 2023), founder of Suzhou InnoLight. Tsinghua Class of 1989, the optical-technology brain of the company; ran the module business as GM from 2017 and took the chair in 2023 ``. Track record: built the world's #1 optical-module franchise from a Tsinghua spin-out into a ~$5.6bn-revenue leader — a genuinely strong, quantified operating record. Archetype: technical founder-operator now running the combined entity.
• Wang Weixiu — founder/controller, now Honorary Chairman. The switchgear entrepreneur who bought the optical business in 2016 and — notably — handed operational control to Liu Sheng (GM in 2017, Chairman in 2023) rather than clinging to it ``. That capital-allocation move (recognise the better operator, buy the asset, step back) is the best thing in the governance file.
• Skin in the game: Wang Weixiu and family hold ~17.41% — a fortune of ~¥196.7bn at the highs ``. Founder ownership is high and aligned.
• Capital allocation: reinvesting almost everything into capacity (capex/capacity +34% YoY) and high-speed R&D; dividend is a token. For a hypergrowth supply-constrained business this is the right call. ROE TTM ~54% `` is exceptional (cycle-flattered, but real).
• Red flags (do not skip — see Lens 10 for the regulatory specifics): (1) a 2021 CSRC Shandong Bureau warning letter against then-controller Wang Xiaodong over an unfulfilled share-pledge commitment from the 2017 acquisition ; (2) **controlling-shareholder block-trade selling: Zhongji Investment cut ~5.5m shares Nov 2025–Jan 2026 at avg ¥521.73, ~¥2.87bn cashed out** — insiders monetising into the AI rally is a yellow flag on top-tick risk, even if the stock kept rising after. Net: competent, aligned, founder-led — with A-share-governance caveats (pledge history, opportunistic insider selling) that warrant ongoing monitoring rather than a red light.

Lens 10 · Forensic Red Flags

Accounting/quality risks (web-only — flagged as low-visibility; a filings-grounded refresh is required to close these):

• Revenue concentration & recognition: >90% overseas, >50% Nvidia wallet — concentration risk is extreme even if recognition is clean. No evidence of channel-stuffing, but the customer-prebooking dynamic ("all of 2026 ordered") is the kind of thing to verify against shipment/revenue timing in filings.
• Inventory & receivables (the #1 thing the web can't show): a module maker hoarding scarce EML/DSP to secure supply can build inventory ahead of revenue; hyperscaler payment terms can stretch receivables. Unverifiable web-only — explicit open item.
• Margin sustainability: 42.6% GM and the FY26 step-up are mix-driven; if 1.6T pricing or EML cost moves against them, the non-cash-flattered portion of the story compresses fast. No SBC-flatters-non-GAAP issue of the US kind here (A-share reporting), but watch capitalised R&D and government grants common to Chinese champions.
• Cash flow vs. earnings: not cleanly sourced web-only — n/a, flag for filings refresh.

Regulatory findings (required sub-section):

• SEC (EDGAR LR/AAER): none possible — Zhongji Innolight has no CIK and does not file with the SEC; regulatory/regulatory-findings.md (generated 2026-06-20) confirms total_sec_findings: 0 and notes the no-CIK limitation ``.
• CSRC (the relevant regulator): Yes — one finding. On 6 Dec 2021 the CSRC Shandong Bureau issued a warning letter to controller Wang Xiaodong for failing to honour a share-pledge-ratio commitment tied to the 2017 InnoLight acquisition (unpledged shares fell >5% below Liu Sheng's from Sep 2017–Jul 2020); recorded in the securities-market integrity file ``. A governance/administrative measure, not fraud — but a real regulator-on-record item.
• Non-SEC (FTC/DOJ/FDA/etc.): web search surfaced no material US enforcement action against Innolight. (US risk for this name is trade/export policy — tariffs, potential Entity-List exposure — which it pre-empts via Thailand/Mexico manufacturing, not an enforcement action.)
• 10-K Item 3: n/a — no SEC filer; no equivalent pulled web-only.
• Summary: No US securities-enforcement findings (no CIK; verified via SEC EDGAR EFTS LR/AAER 2026-06-20). One CSRC Shandong Bureau warning letter (2021, share-pledge commitment) on the controller. Insider block-trade selling Nov 2025–Jan 2026. No US agency enforcement found via web search as of 2026-06-20. Inventory/receivables/cash-flow quality unverifiable web-only — flagged for a filings-grounded refresh.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (EPS, next 3 fiscal years)

No forecast.ts create is logged (per --watchlist rules). All figures built bottom-up from FY25 actuals + the FY26 booking signal; shown as **net-profit paths** (EPS is noisy on an A-share share count not cleanly sourced web-only — ~1.1bn shares implied by the ¥1.111bn dividend at ¥10/10sh).

Anchor: FY25 net profit ¥10.8bn; Q1'26 already ¥5.73bn ``.

• FY2026 (base): Q1'26 annualised naively = ¥22.9bn, but mix/seasonality and 1.6T ramp argue for back-half strength. Base net profit ¥22–26bn (~+110–140% YoY) — driven by full-year 800G + 1.6T ramp at rich mix, customers booked. ``
• Bull ¥28–30bn: 1.6T ramps faster, EML supply loosens enough to fill capacity, margins hold >30% net.
• Bear ¥18–20bn: EML shortage caps unit growth and/or 1.6T pricing comes in soft / Nvidia wallet-share dilution begins → revenue and margin both undershoot.
• FY2027 (base): ¥26–32bn. Growth decelerates as the law of large numbers and CPO/in-house erosion begin to bite at the margin; 1.6T is the volume engine, 3.2T still nascent. ``
• Bull ¥36–40bn (CPO stays 2028+, Innolight keeps >50% 1.6T share); Bear ¥18–24bn (architecture shift + share loss + price erosion → first YoY decline possible).
• FY2028 (base): ¥28–34bn — the fork year. This is where CPO "large-scale 2028–2030" and Nvidia in-house optics either start disintermediating module assembly (bear: flat-to-down) or Innolight successfully migrates up into optical-engine/SiPh content and holds (bull: ¥40bn+). ``

The honest summary: FY26 is almost-locked upside (bookings + backlog); FY27–28 is where the multiple gets decided, and it hinges on (a) EML supply, (b) 1.6T pricing/share, (c) the CPO/in-house timeline. The ~29× forward P/E is cheap if FY26 lands and FY27 holds; expensive if FY27 is the peak.

Lens 12 · Bull vs Bear

Bull case. Innolight is the default volume supplier for the single biggest capex wave in tech history, with the #1 share, the best margins in the group (~20–22% net, GM 42.6% and rising), >50% of Nvidia's wallet, and a first-mover 1.6T qualification — and it trades at the lowest forward P/E in its peer set (~29×) . Customers have booked all of 2026 and are placing 2027 orders; the bottleneck is the industry's ability to *deliver*, and Innolight has the capacity (utilisation 85%, expanding 34%/yr) . If pluggable optics remain the answer through the 1.6T cycle and CPO stays a 2028+ story (as management asserts), earnings compound for two more years and the cheap multiple re-rates. Secular tailwind: optical-component TAM seen reaching $90bn by 2030 ``.

Bear case (permanent-impairment risks).

1. Architecture obsolescence — CPO + in-house. If co-packaged optics and Nvidia's own optical engine move faster than 2028, the module-assembly step Innolight owns gets disintermediated. Nvidia's 2 March 2026 $4bn into Coherent + Lumentum explicitly funds CPO capacity at Innolight's suppliers, and Nvidia is expected to shift volume to internally-developed modules, diluting Innolight's wallet share even as total volume grows ``. This is the thesis-killer.
2. Component squeeze. Innolight makes neither the DSP (Broadcom/Marvell) nor the EML (Lumentum/Coherent); it is a margin-taker between two concentrated US supplier bases and a concentrated US customer base. EML is a physical shortage that caps unit growth ``.
3. Concentration + geopolitics. >90% overseas, one end-demand (AI capex), a few customers, a Chinese company exposed to US trade policy (tariffs/Entity-List), hedged only by an overseas-manufacturing ramp that must execute flawlessly.

Pre-mortem (18 months out, thesis broke). Most likely failure: EML supply caps 1.6T volume while Nvidia simultaneously dual-sources harder and accelerates LPO/in-house, so Innolight's revenue grows but share and price slip — FY27 net profit is flat YoY, the "growth" narrative cracks, and a 29× forward multiple on a no-growth cyclical re-rates to the mid-teens. A capex digestion (post-DeepSeek-style efficiency scare) would amplify it.

Are multiples too high? On forward earnings, no — it's the cheapest in the group. On normalized/through-cycle earnings, the market clearly thinks the risk is real (hence the discount to Coherent/Lumentum/Eoptolink). The multiple isn't the danger; the durability of the earnings is.

Contrarian view (what the market refuses to see): Bears fixate on CPO killing modules; the underappreciated path is that Innolight migrates into the optical engine / silicon-photonics content (it's an LPO MSA founder and is developing SiPh) and ends up a supplier into CPO systems rather than a casualty of them — keeping content-per-GPU even as the form factor changes. If that's right, the architecture transition is a re-mix, not a death, and the cheap multiple is a gift. The other underappreciated point: at ~29× forward with >50% ROE and a booked year, a lot of bad news is already in the price.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case.

• What structurally breaks the model: Innolight's entire economic role is assembling other people's silicon into a pluggable box. CPO removes the box; in-house removes the assembler; LPO removes the DSP content. Every leading-edge architecture roadmap subtracts value from the module-assembly node. The bull's "they'll migrate into optical engines" is a hope, not a demonstrated, qualified, margin-accretive business yet.
• Revenue concentration: >50% of the growth engine is one customer's (Nvidia's) wallet, and that customer is publicly building the thing that disintermediates you ($4bn CPO, in-house modules) and dual-sourcing to Eoptolink. The concentration isn't just risky — it's concentrated in your most motivated disruptor.
• Why the moat is weaker than bulls think: the qual lead resets every generation; the cost edge invites Nvidia to commoditise you on purpose (qualify more Chinese suppliers, push price); and you have zero pricing power on your two most expensive inputs. A "moat" you re-dig every 18 months against a customer who profits from filling it in is not a moat.
• Most dangerous competitor bulls underestimate: not Eoptolink — Broadcom. Broadcom owns the DSP and is shipping 50,000+ CPO switches by 2025 ``; it can collapse the module layer into the switch and capture the economics directly. Innolight cannot answer that with assembly scale.
• Worst capital-allocation / governance: the insider block-selling (¥2.87bn, Nov'25–Jan'26) into the rally and the **2021 CSRC pledge warning** — small, but exactly the texture a short cites for "insiders know the cycle peaks."
• What must hold for today's price: pluggables survive the 1.6T cycle; EML supply loosens enough to grow units; Nvidia wallet-share erosion is gradual not abrupt; FY26 backlog converts to margin not just revenue. If growth disappoints 20–30% (FY27 net profit flat-to-down), a 29× forward multiple on a now-cyclical re-rates toward 12–15× → 40–55% downside even before sentiment overshoot.
• Single permanent-impairment scenario & plausibility: Nvidia + hyperscalers standardise on CPO/in-house for the 1.6T→3.2T transition by 2027–2028 and Innolight fails to win qualified optical-engine content → it becomes a low-margin legacy-speed assembler. Plausibility: moderate — management says CSPs have no large-scale CPO plan for 2026–27, but the capital (Nvidia's $4bn) is already committed, and roadmaps move faster than incumbents admit.

Lens 14 · Management Questions (ordered by information value)

1. CPO/in-house timeline: Beyond "no large-scale 2026–27 CSP deployment," what specific qualified design wins do you have for optical engines / SiPh content inside CPO systems — i.e. what is your content-per-GPU in a co-packaged world, and is it accretive or dilutive to today's module margin?
2. Nvidia wallet share: As Nvidia ramps internally-developed modules, what is your forecast trajectory of Nvidia wallet share through the 1.6T and 3.2T transitions, and what offsets it (new hyperscaler logos, non-Nvidia accelerators)?
3. EML supply: Is leading-edge EML the binding constraint on your 2026–27 unit volume? What share of your EML is single-sourced from Lumentum/Coherent, and what is your qualified Chinese-EML roadmap (Source Photonics, in-house)?
4. 1.6T pricing & margin: What is the gross-margin profile of 1.6T vs. 800G at maturity, and how much of FY26's margin is mix (durable) vs. shortage-driven pricing (transient)?
5. Inventory & receivables: Given "customers booked all of 2026," how much component inventory have you pre-committed, and how are receivables/payment terms trending with hyperscaler customers?
6. DSP dependency: What is your strategy as a price-taker on Broadcom/Marvell DSP — does LPO (DSP-less) materially change your BOM economics and supplier power, and what % of your 800G/1.6T mix is LPO?
7. 3.2T: What is your 3.2T qualification timeline and is the architecture pluggable, LPO, or co-packaged — i.e. do you own the 3.2T node?
8. Geographic/geopolitical: What % of capacity is in Thailand/Mexico vs. China today, what's the FY27 target, and what is your contingency if US tariffs or Entity-List action hit Chinese optical suppliers?
9. Capacity discipline: With utilisation at 85% and capacity +34%/yr, how do you avoid over-building into a potential capex digestion? What's your capacity plan if AI capex growth halves?
10. Customer power: As Nvidia/hyperscalers qualify more second sources, what is your defence against deliberate commoditisation — what keeps you from becoming a price-taker on output as well as input?
11. Capital allocation: With ~54% ROE and a token dividend, how should investors think about reinvestment runway vs. returns over the next 3 years — at what point does capital intensity outrun the opportunity?
12. Insider selling: What should investors read into the controlling-shareholder block sales of ¥2.87bn in Nov 2025–Jan 2026?
13. Governance: What has changed in share-pledge and related-party governance since the 2021 CSRC warning letter?
14. Competitive intel: Who do you regard as the most dangerous competitor over a 3-year horizon — Eoptolink, Broadcom (CPO), or Nvidia in-house — and why?
15. The fork: If you had to name the single scenario that permanently impairs this business, what is it, and what are you doing today to insure against it?

---